Improving hail nowcasting using multi-source data and a WMHDA-enhanced DIFF-Transformer

HailPre is a deep learning framework developed for accurate 60-minute hail nowcasting at 10-minute intervals, tailored for regions with complex terrain and imbalanced hail event distributions. It integrates multi-source remote sensing data, including weather radar mosaics, Himawari-8/9 satellite infrared imagery, lightning strike records, gridded precipitation, topographic elevation, and land cover, fused with expert-verified ground-based hail observations from 2022–2024 in Guizhou. The framework employs a DIFF-Transformer architecture, enhanced with a Weighted Multi-Head Differential Attention (WMHDA) mechanism, to capture complex spatiotemporal patterns in hail nowcasting. A composite loss function, combining binary cross-entropy, Dice, and focal losses, addresses class imbalance by emphasizing rare hail events. Ablation experiments on the model confirm the contribution of the WMHDA module. Experimental evaluation demonstrates HailPre’s superior performance, achieving an average 60-minute CSI (Critical Success Index) of 0.4617 and POD (Probability of Detection) of 0.6406, outperforming baseline models by 6%–39% in neighborhood-based metrics. Ablation studies on data sources and SHAP (SHapley Additive exPlanations) analysis reveal that satellite-derived predictors play the dominant role, followed by static and radar data. These findings highlight HailPre’s ability to leverage diverse data and specialized architecture for robust hail nowcasting, offering practical value for operational weather forecasting and disaster preparedness in hail-prone regions. HailPre模型旨在解决复杂地形下的类不平衡冰雹临近预报问题.以人工标注的冰雹观测为真值, 多源数据作为输入, 模型采用DIFF-Transformer结合加权多头差分注意力机制 (WMHDA) 作为编码器, 有效建模多源时空关系及长时依赖.复合损失函数增强了对冰雹稀疏特征的适应性, 模块消融实验证实WMHDA在冰雹临近预报中的贡献.HailPre在测试集上的邻域CSI和POD分别达到0.4617和0.6406, 较对比模型提升6–39%, 且表现稳定.数据源消融实验和SHAP分析显示, 卫星数据贡献最大, 其次为静态和雷达数据.实验结果表明, 多源融合与WMHDA机制均显著提升了冰雹预报性能.